xref: /openbsd-src/sys/uvm/uvm_pdaemon.c (revision 7350f337b9e3eb4461d99580e625c7ef148d107c)
1 /*	$OpenBSD: uvm_pdaemon.c,v 1.82 2019/05/10 02:33:57 beck Exp $	*/
2 /*	$NetBSD: uvm_pdaemon.c,v 1.23 2000/08/20 10:24:14 bjh21 Exp $	*/
3 
4 /*
5  * Copyright (c) 1997 Charles D. Cranor and Washington University.
6  * Copyright (c) 1991, 1993, The Regents of the University of California.
7  *
8  * All rights reserved.
9  *
10  * This code is derived from software contributed to Berkeley by
11  * The Mach Operating System project at Carnegie-Mellon University.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. Neither the name of the University nor the names of its contributors
22  *    may be used to endorse or promote products derived from this software
23  *    without specific prior written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
27  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
28  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
29  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
30  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
31  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
33  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
34  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35  * SUCH DAMAGE.
36  *
37  *	@(#)vm_pageout.c        8.5 (Berkeley) 2/14/94
38  * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp
39  *
40  *
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  */
64 
65 /*
66  * uvm_pdaemon.c: the page daemon
67  */
68 
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/kernel.h>
72 #include <sys/pool.h>
73 #include <sys/proc.h>
74 #include <sys/buf.h>
75 #include <sys/mount.h>
76 #include <sys/atomic.h>
77 
78 #ifdef HIBERNATE
79 #include <sys/hibernate.h>
80 #endif
81 
82 #include <uvm/uvm.h>
83 
84 /*
85  * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate
86  * in a pass thru the inactive list when swap is full.  the value should be
87  * "small"... if it's too large we'll cycle the active pages thru the inactive
88  * queue too quickly to for them to be referenced and avoid being freed.
89  */
90 
91 #define UVMPD_NUMDIRTYREACTS 16
92 
93 
94 /*
95  * local prototypes
96  */
97 
98 void		uvmpd_scan(void);
99 boolean_t	uvmpd_scan_inactive(struct pglist *);
100 void		uvmpd_tune(void);
101 void		uvmpd_drop(struct pglist *);
102 
103 /*
104  * uvm_wait: wait (sleep) for the page daemon to free some pages
105  *
106  * => should be called with all locks released
107  * => should _not_ be called by the page daemon (to avoid deadlock)
108  */
109 
110 void
111 uvm_wait(const char *wmsg)
112 {
113 	int	timo = 0;
114 
115 #ifdef DIAGNOSTIC
116 	if (curproc == &proc0)
117 		panic("%s: cannot sleep for memory during boot", __func__);
118 #endif
119 
120 	/* check for page daemon going to sleep (waiting for itself) */
121 	if (curproc == uvm.pagedaemon_proc) {
122 		printf("uvm_wait emergency bufbackoff\n");
123 		if (bufbackoff(NULL, 4) == 0)
124 			return;
125 		/*
126 		 * now we have a problem: the pagedaemon wants to go to
127 		 * sleep until it frees more memory.   but how can it
128 		 * free more memory if it is asleep?  that is a deadlock.
129 		 * we have two options:
130 		 *  [1] panic now
131 		 *  [2] put a timeout on the sleep, thus causing the
132 		 *      pagedaemon to only pause (rather than sleep forever)
133 		 *
134 		 * note that option [2] will only help us if we get lucky
135 		 * and some other process on the system breaks the deadlock
136 		 * by exiting or freeing memory (thus allowing the pagedaemon
137 		 * to continue).  for now we panic if DEBUG is defined,
138 		 * otherwise we hope for the best with option [2] (better
139 		 * yet, this should never happen in the first place!).
140 		 */
141 
142 		printf("pagedaemon: deadlock detected!\n");
143 		timo = hz >> 3;		/* set timeout */
144 #if defined(DEBUG)
145 		/* DEBUG: panic so we can debug it */
146 		panic("pagedaemon deadlock");
147 #endif
148 	}
149 
150 	uvm_lock_fpageq();
151 	wakeup(&uvm.pagedaemon);		/* wake the daemon! */
152 	msleep(&uvmexp.free, &uvm.fpageqlock, PVM | PNORELOCK, wmsg, timo);
153 }
154 
155 /*
156  * uvmpd_tune: tune paging parameters
157  *
158  * => called whenever memory is added to (or removed from?) the system
159  * => caller must call with page queues locked
160  */
161 
162 void
163 uvmpd_tune(void)
164 {
165 
166 	uvmexp.freemin = uvmexp.npages / 30;
167 
168 	/* between 16k and 512k */
169 	/* XXX:  what are these values good for? */
170 	uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT);
171 #if 0
172 	uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT);
173 #endif
174 
175 	/* Make sure there's always a user page free. */
176 	if (uvmexp.freemin < uvmexp.reserve_kernel + 1)
177 		uvmexp.freemin = uvmexp.reserve_kernel + 1;
178 
179 	uvmexp.freetarg = (uvmexp.freemin * 4) / 3;
180 	if (uvmexp.freetarg <= uvmexp.freemin)
181 		uvmexp.freetarg = uvmexp.freemin + 1;
182 
183 	/* uvmexp.inactarg: computed in main daemon loop */
184 
185 	uvmexp.wiredmax = uvmexp.npages / 3;
186 }
187 
188 /*
189  * Indicate to the page daemon that a nowait call failed and it should
190  * recover at least some memory in the most restricted region (assumed
191  * to be dma_constraint).
192  */
193 volatile int uvm_nowait_failed;
194 
195 /*
196  * uvm_pageout: the main loop for the pagedaemon
197  */
198 void
199 uvm_pageout(void *arg)
200 {
201 	struct uvm_constraint_range constraint;
202 	struct uvm_pmalloc *pma;
203 	int work_done;
204 	int npages = 0;
205 
206 	/* ensure correct priority and set paging parameters... */
207 	uvm.pagedaemon_proc = curproc;
208 	(void) spl0();
209 	uvm_lock_pageq();
210 	npages = uvmexp.npages;
211 	uvmpd_tune();
212 	uvm_unlock_pageq();
213 
214 	for (;;) {
215 		long size;
216 	  	work_done = 0; /* No work done this iteration. */
217 
218 		uvm_lock_fpageq();
219 		if (!uvm_nowait_failed && TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
220 			msleep(&uvm.pagedaemon, &uvm.fpageqlock, PVM,
221 			    "pgdaemon", 0);
222 			uvmexp.pdwoke++;
223 		}
224 
225 		if ((pma = TAILQ_FIRST(&uvm.pmr_control.allocs)) != NULL) {
226 			pma->pm_flags |= UVM_PMA_BUSY;
227 			constraint = pma->pm_constraint;
228 		} else {
229 			if (uvm_nowait_failed) {
230 				/*
231 				 * XXX realisticly, this is what our
232 				 * nowait callers probably care about
233 				 */
234 				constraint = dma_constraint;
235 				uvm_nowait_failed = 0;
236 			} else
237 				constraint = no_constraint;
238 		}
239 
240 		uvm_unlock_fpageq();
241 
242 		/* now lock page queues and recompute inactive count */
243 		uvm_lock_pageq();
244 		if (npages != uvmexp.npages) {	/* check for new pages? */
245 			npages = uvmexp.npages;
246 			uvmpd_tune();
247 		}
248 
249 		uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3;
250 		if (uvmexp.inactarg <= uvmexp.freetarg) {
251 			uvmexp.inactarg = uvmexp.freetarg + 1;
252 		}
253 
254 		/* Reclaim pages from the buffer cache if possible. */
255 		size = 0;
256 		if (pma != NULL)
257 			size += pma->pm_size >> PAGE_SHIFT;
258 		if (uvmexp.free - BUFPAGES_DEFICIT < uvmexp.freetarg)
259 			size += uvmexp.freetarg - (uvmexp.free -
260 			    BUFPAGES_DEFICIT);
261 		if (size == 0)
262 			size = 16; /* XXX */
263 		uvm_unlock_pageq();
264 		(void) bufbackoff(&constraint, size * 2);
265 		uvm_lock_pageq();
266 
267 		/* Scan if needed to meet our targets. */
268 		if (pma != NULL ||
269 		    ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg) ||
270 		    ((uvmexp.inactive + BUFPAGES_INACT) < uvmexp.inactarg)) {
271 			uvmpd_scan();
272 			work_done = 1; /* XXX we hope... */
273 		}
274 
275 		/*
276 		 * if there's any free memory to be had,
277 		 * wake up any waiters.
278 		 */
279 		uvm_lock_fpageq();
280 		if (uvmexp.free > uvmexp.reserve_kernel ||
281 		    uvmexp.paging == 0) {
282 			wakeup(&uvmexp.free);
283 		}
284 
285 		if (pma != NULL) {
286 			pma->pm_flags &= ~UVM_PMA_BUSY;
287 			if (!work_done)
288 				pma->pm_flags |= UVM_PMA_FAIL;
289 			if (pma->pm_flags & (UVM_PMA_FAIL | UVM_PMA_FREED)) {
290 				pma->pm_flags &= ~UVM_PMA_LINKED;
291 				TAILQ_REMOVE(&uvm.pmr_control.allocs, pma,
292 				    pmq);
293 			}
294 			wakeup(pma);
295 		}
296 		uvm_unlock_fpageq();
297 
298 		/* scan done. unlock page queues (only lock we are holding) */
299 		uvm_unlock_pageq();
300 
301 		sched_pause(yield);
302 	}
303 	/*NOTREACHED*/
304 }
305 
306 
307 /*
308  * uvm_aiodone_daemon:  main loop for the aiodone daemon.
309  */
310 void
311 uvm_aiodone_daemon(void *arg)
312 {
313 	int s, free;
314 	struct buf *bp, *nbp;
315 
316 	uvm.aiodoned_proc = curproc;
317 
318 	for (;;) {
319 		/*
320 		 * Check for done aio structures. If we've got structures to
321 		 * process, do so. Otherwise sleep while avoiding races.
322 		 */
323 		mtx_enter(&uvm.aiodoned_lock);
324 		while ((bp = TAILQ_FIRST(&uvm.aio_done)) == NULL)
325 			msleep(&uvm.aiodoned, &uvm.aiodoned_lock,
326 			    PVM, "aiodoned", 0);
327 		/* Take the list for ourselves. */
328 		TAILQ_INIT(&uvm.aio_done);
329 		mtx_leave(&uvm.aiodoned_lock);
330 
331 		/* process each i/o that's done. */
332 		free = uvmexp.free;
333 		while (bp != NULL) {
334 			if (bp->b_flags & B_PDAEMON) {
335 				uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT;
336 			}
337 			nbp = TAILQ_NEXT(bp, b_freelist);
338 			s = splbio();	/* b_iodone must by called at splbio */
339 			(*bp->b_iodone)(bp);
340 			splx(s);
341 			bp = nbp;
342 
343 			sched_pause(yield);
344 		}
345 		uvm_lock_fpageq();
346 		wakeup(free <= uvmexp.reserve_kernel ? &uvm.pagedaemon :
347 		    &uvmexp.free);
348 		uvm_unlock_fpageq();
349 	}
350 }
351 
352 
353 
354 /*
355  * uvmpd_scan_inactive: scan an inactive list for pages to clean or free.
356  *
357  * => called with page queues locked
358  * => we work on meeting our free target by converting inactive pages
359  *    into free pages.
360  * => we handle the building of swap-backed clusters
361  * => we return TRUE if we are exiting because we met our target
362  */
363 
364 boolean_t
365 uvmpd_scan_inactive(struct pglist *pglst)
366 {
367 	boolean_t retval = FALSE;	/* assume we haven't hit target */
368 	int free, result;
369 	struct vm_page *p, *nextpg;
370 	struct uvm_object *uobj;
371 	struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp;
372 	int npages;
373 	struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; 	/* XXX: see below */
374 	int swnpages, swcpages;				/* XXX: see below */
375 	int swslot;
376 	struct vm_anon *anon;
377 	boolean_t swap_backed;
378 	vaddr_t start;
379 	int dirtyreacts;
380 
381 	/*
382 	 * note: we currently keep swap-backed pages on a separate inactive
383 	 * list from object-backed pages.   however, merging the two lists
384 	 * back together again hasn't been ruled out.   thus, we keep our
385 	 * swap cluster in "swpps" rather than in pps (allows us to mix
386 	 * clustering types in the event of a mixed inactive queue).
387 	 */
388 	/*
389 	 * swslot is non-zero if we are building a swap cluster.  we want
390 	 * to stay in the loop while we have a page to scan or we have
391 	 * a swap-cluster to build.
392 	 */
393 	swslot = 0;
394 	swnpages = swcpages = 0;
395 	free = 0;
396 	dirtyreacts = 0;
397 
398 	for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) {
399 		/*
400 		 * note that p can be NULL iff we have traversed the whole
401 		 * list and need to do one final swap-backed clustered pageout.
402 		 */
403 		uobj = NULL;
404 		anon = NULL;
405 
406 		if (p) {
407 			/*
408 			 * update our copy of "free" and see if we've met
409 			 * our target
410 			 */
411 			free = uvmexp.free - BUFPAGES_DEFICIT;
412 
413 			if (free + uvmexp.paging >= uvmexp.freetarg << 2 ||
414 			    dirtyreacts == UVMPD_NUMDIRTYREACTS) {
415 				retval = TRUE;
416 
417 				if (swslot == 0) {
418 					/* exit now if no swap-i/o pending */
419 					break;
420 				}
421 
422 				/* set p to null to signal final swap i/o */
423 				p = NULL;
424 			}
425 		}
426 
427 		if (p) {	/* if (we have a new page to consider) */
428 			/*
429 			 * we are below target and have a new page to consider.
430 			 */
431 			uvmexp.pdscans++;
432 			nextpg = TAILQ_NEXT(p, pageq);
433 
434 			/*
435 			 * move referenced pages back to active queue and
436 			 * skip to next page (unlikely to happen since
437 			 * inactive pages shouldn't have any valid mappings
438 			 * and we cleared reference before deactivating).
439 			 */
440 
441 			if (pmap_is_referenced(p)) {
442 				uvm_pageactivate(p);
443 				uvmexp.pdreact++;
444 				continue;
445 			}
446 
447 			if (p->pg_flags & PQ_ANON) {
448 				anon = p->uanon;
449 				KASSERT(anon != NULL);
450 				if (p->pg_flags & PG_BUSY) {
451 					uvmexp.pdbusy++;
452 					/* someone else owns page, skip it */
453 					continue;
454 				}
455 				uvmexp.pdanscan++;
456 			} else {
457 				uobj = p->uobject;
458 				KASSERT(uobj != NULL);
459 				if (p->pg_flags & PG_BUSY) {
460 					uvmexp.pdbusy++;
461 					/* someone else owns page, skip it */
462 					continue;
463 				}
464 				uvmexp.pdobscan++;
465 			}
466 
467 			/*
468 			 * we now have the page queues locked.
469 			 * the page is not busy.   if the page is clean we
470 			 * can free it now and continue.
471 			 */
472 			if (p->pg_flags & PG_CLEAN) {
473 				if (p->pg_flags & PQ_SWAPBACKED) {
474 					/* this page now lives only in swap */
475 					uvmexp.swpgonly++;
476 				}
477 
478 				/* zap all mappings with pmap_page_protect... */
479 				pmap_page_protect(p, PROT_NONE);
480 				uvm_pagefree(p);
481 				uvmexp.pdfreed++;
482 
483 				if (anon) {
484 
485 					/*
486 					 * an anonymous page can only be clean
487 					 * if it has backing store assigned.
488 					 */
489 
490 					KASSERT(anon->an_swslot != 0);
491 
492 					/* remove from object */
493 					anon->an_page = NULL;
494 				}
495 				continue;
496 			}
497 
498 			/*
499 			 * this page is dirty, skip it if we'll have met our
500 			 * free target when all the current pageouts complete.
501 			 */
502 			if (free + uvmexp.paging > uvmexp.freetarg << 2) {
503 				continue;
504 			}
505 
506 			/*
507 			 * this page is dirty, but we can't page it out
508 			 * since all pages in swap are only in swap.
509 			 * reactivate it so that we eventually cycle
510 			 * all pages thru the inactive queue.
511 			 */
512 			KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
513 			if ((p->pg_flags & PQ_SWAPBACKED) &&
514 			    uvmexp.swpgonly == uvmexp.swpages) {
515 				dirtyreacts++;
516 				uvm_pageactivate(p);
517 				continue;
518 			}
519 
520 			/*
521 			 * if the page is swap-backed and dirty and swap space
522 			 * is full, free any swap allocated to the page
523 			 * so that other pages can be paged out.
524 			 */
525 			KASSERT(uvmexp.swpginuse <= uvmexp.swpages);
526 			if ((p->pg_flags & PQ_SWAPBACKED) &&
527 			    uvmexp.swpginuse == uvmexp.swpages) {
528 
529 				if ((p->pg_flags & PQ_ANON) &&
530 				    p->uanon->an_swslot) {
531 					uvm_swap_free(p->uanon->an_swslot, 1);
532 					p->uanon->an_swslot = 0;
533 				}
534 				if (p->pg_flags & PQ_AOBJ) {
535 					uao_dropswap(p->uobject,
536 						     p->offset >> PAGE_SHIFT);
537 				}
538 			}
539 
540 			/*
541 			 * the page we are looking at is dirty.   we must
542 			 * clean it before it can be freed.  to do this we
543 			 * first mark the page busy so that no one else will
544 			 * touch the page.   we write protect all the mappings
545 			 * of the page so that no one touches it while it is
546 			 * in I/O.
547 			 */
548 
549 			swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0);
550 			atomic_setbits_int(&p->pg_flags, PG_BUSY);
551 			UVM_PAGE_OWN(p, "scan_inactive");
552 			pmap_page_protect(p, PROT_READ);
553 			uvmexp.pgswapout++;
554 
555 			/*
556 			 * for swap-backed pages we need to (re)allocate
557 			 * swap space.
558 			 */
559 			if (swap_backed) {
560 				/* free old swap slot (if any) */
561 				if (anon) {
562 					if (anon->an_swslot) {
563 						uvm_swap_free(anon->an_swslot,
564 						    1);
565 						anon->an_swslot = 0;
566 					}
567 				} else {
568 					uao_dropswap(uobj,
569 						     p->offset >> PAGE_SHIFT);
570 				}
571 
572 				/* start new cluster (if necessary) */
573 				if (swslot == 0) {
574 					swnpages = MAXBSIZE >> PAGE_SHIFT;
575 					swslot = uvm_swap_alloc(&swnpages,
576 					    TRUE);
577 					if (swslot == 0) {
578 						/* no swap?  give up! */
579 						atomic_clearbits_int(
580 						    &p->pg_flags,
581 						    PG_BUSY);
582 						UVM_PAGE_OWN(p, NULL);
583 						continue;
584 					}
585 					swcpages = 0;	/* cluster is empty */
586 				}
587 
588 				/* add block to cluster */
589 				swpps[swcpages] = p;
590 				if (anon)
591 					anon->an_swslot = swslot + swcpages;
592 				else
593 					uao_set_swslot(uobj,
594 					    p->offset >> PAGE_SHIFT,
595 					    swslot + swcpages);
596 				swcpages++;
597 			}
598 		} else {
599 			/* if p == NULL we must be doing a last swap i/o */
600 			swap_backed = TRUE;
601 		}
602 
603 		/*
604 		 * now consider doing the pageout.
605 		 *
606 		 * for swap-backed pages, we do the pageout if we have either
607 		 * filled the cluster (in which case (swnpages == swcpages) or
608 		 * run out of pages (p == NULL).
609 		 *
610 		 * for object pages, we always do the pageout.
611 		 */
612 		if (swap_backed) {
613 			if (p) {	/* if we just added a page to cluster */
614 				/* cluster not full yet? */
615 				if (swcpages < swnpages)
616 					continue;
617 			}
618 
619 			/* starting I/O now... set up for it */
620 			npages = swcpages;
621 			ppsp = swpps;
622 			/* for swap-backed pages only */
623 			start = (vaddr_t) swslot;
624 
625 			/* if this is final pageout we could have a few
626 			 * extra swap blocks */
627 			if (swcpages < swnpages) {
628 				uvm_swap_free(swslot + swcpages,
629 				    (swnpages - swcpages));
630 			}
631 		} else {
632 			/* normal object pageout */
633 			ppsp = pps;
634 			npages = sizeof(pps) / sizeof(struct vm_page *);
635 			/* not looked at because PGO_ALLPAGES is set */
636 			start = 0;
637 		}
638 
639 		/*
640 		 * now do the pageout.
641 		 *
642 		 * for swap_backed pages we have already built the cluster.
643 		 * for !swap_backed pages, uvm_pager_put will call the object's
644 		 * "make put cluster" function to build a cluster on our behalf.
645 		 *
646 		 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct
647 		 * it to free the cluster pages for us on a successful I/O (it
648 		 * always does this for un-successful I/O requests).  this
649 		 * allows us to do clustered pageout without having to deal
650 		 * with cluster pages at this level.
651 		 *
652 		 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST:
653 		 *  IN: locked: page queues
654 		 * OUT: locked:
655 		 *     !locked: pageqs
656 		 */
657 
658 		uvmexp.pdpageouts++;
659 		result = uvm_pager_put(swap_backed ? NULL : uobj, p,
660 		    &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0);
661 
662 		/*
663 		 * if we did i/o to swap, zero swslot to indicate that we are
664 		 * no longer building a swap-backed cluster.
665 		 */
666 
667 		if (swap_backed)
668 			swslot = 0;		/* done with this cluster */
669 
670 		/*
671 		 * first, we check for VM_PAGER_PEND which means that the
672 		 * async I/O is in progress and the async I/O done routine
673 		 * will clean up after us.   in this case we move on to the
674 		 * next page.
675 		 *
676 		 * there is a very remote chance that the pending async i/o can
677 		 * finish _before_ we get here.   if that happens, our page "p"
678 		 * may no longer be on the inactive queue.   so we verify this
679 		 * when determining the next page (starting over at the head if
680 		 * we've lost our inactive page).
681 		 */
682 
683 		if (result == VM_PAGER_PEND) {
684 			uvmexp.paging += npages;
685 			uvm_lock_pageq();
686 			uvmexp.pdpending++;
687 			if (p) {
688 				if (p->pg_flags & PQ_INACTIVE)
689 					nextpg = TAILQ_NEXT(p, pageq);
690 				else
691 					nextpg = TAILQ_FIRST(pglst);
692 			} else {
693 				nextpg = NULL;
694 			}
695 			continue;
696 		}
697 
698 		/* clean up "p" if we have one */
699 		if (p) {
700 			/*
701 			 * the I/O request to "p" is done and uvm_pager_put
702 			 * has freed any cluster pages it may have allocated
703 			 * during I/O.  all that is left for us to do is
704 			 * clean up page "p" (which is still PG_BUSY).
705 			 *
706 			 * our result could be one of the following:
707 			 *   VM_PAGER_OK: successful pageout
708 			 *
709 			 *   VM_PAGER_AGAIN: tmp resource shortage, we skip
710 			 *     to next page
711 			 *   VM_PAGER_{FAIL,ERROR,BAD}: an error.   we
712 			 *     "reactivate" page to get it out of the way (it
713 			 *     will eventually drift back into the inactive
714 			 *     queue for a retry).
715 			 *   VM_PAGER_UNLOCK: should never see this as it is
716 			 *     only valid for "get" operations
717 			 */
718 
719 			/* relock p's object: page queues not lock yet, so
720 			 * no need for "try" */
721 
722 #ifdef DIAGNOSTIC
723 			if (result == VM_PAGER_UNLOCK)
724 				panic("pagedaemon: pageout returned "
725 				    "invalid 'unlock' code");
726 #endif
727 
728 			/* handle PG_WANTED now */
729 			if (p->pg_flags & PG_WANTED)
730 				wakeup(p);
731 
732 			atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED);
733 			UVM_PAGE_OWN(p, NULL);
734 
735 			/* released during I/O? Can only happen for anons */
736 			if (p->pg_flags & PG_RELEASED) {
737 				KASSERT(anon != NULL);
738 				/*
739 				 * remove page so we can get nextpg,
740 				 * also zero out anon so we don't use
741 				 * it after the free.
742 				 */
743 				anon->an_page = NULL;
744 				p->uanon = NULL;
745 
746 				uvm_anfree(anon);	/* kills anon */
747 				pmap_page_protect(p, PROT_NONE);
748 				anon = NULL;
749 				uvm_lock_pageq();
750 				nextpg = TAILQ_NEXT(p, pageq);
751 				/* free released page */
752 				uvm_pagefree(p);
753 			} else {	/* page was not released during I/O */
754 				uvm_lock_pageq();
755 				nextpg = TAILQ_NEXT(p, pageq);
756 				if (result != VM_PAGER_OK) {
757 					/* pageout was a failure... */
758 					if (result != VM_PAGER_AGAIN)
759 						uvm_pageactivate(p);
760 					pmap_clear_reference(p);
761 					/* XXXCDC: if (swap_backed) FREE p's
762 					 * swap block? */
763 				} else {
764 					/* pageout was a success... */
765 					pmap_clear_reference(p);
766 					pmap_clear_modify(p);
767 					atomic_setbits_int(&p->pg_flags,
768 					    PG_CLEAN);
769 				}
770 			}
771 
772 			/*
773 			 * drop object lock (if there is an object left).   do
774 			 * a safety check of nextpg to make sure it is on the
775 			 * inactive queue (it should be since PG_BUSY pages on
776 			 * the inactive queue can't be re-queued [note: not
777 			 * true for active queue]).
778 			 */
779 
780 			if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) {
781 				nextpg = TAILQ_FIRST(pglst);	/* reload! */
782 			}
783 		} else {
784 			/*
785 			 * if p is null in this loop, make sure it stays null
786 			 * in the next loop.
787 			 */
788 			nextpg = NULL;
789 
790 			/*
791 			 * lock page queues here just so they're always locked
792 			 * at the end of the loop.
793 			 */
794 			uvm_lock_pageq();
795 		}
796 	}
797 	return (retval);
798 }
799 
800 /*
801  * uvmpd_scan: scan the page queues and attempt to meet our targets.
802  *
803  * => called with pageq's locked
804  */
805 
806 void
807 uvmpd_scan(void)
808 {
809 	int free, inactive_shortage, swap_shortage, pages_freed;
810 	struct vm_page *p, *nextpg;
811 	struct uvm_object *uobj;
812 	boolean_t got_it;
813 
814 	uvmexp.pdrevs++;		/* counter */
815 	uobj = NULL;
816 
817 	/*
818 	 * get current "free" page count
819 	 */
820 	free = uvmexp.free - BUFPAGES_DEFICIT;
821 
822 #ifndef __SWAP_BROKEN
823 	/*
824 	 * swap out some processes if we are below our free target.
825 	 * we need to unlock the page queues for this.
826 	 */
827 	if (free < uvmexp.freetarg) {
828 		uvmexp.pdswout++;
829 		uvm_unlock_pageq();
830 		uvm_swapout_threads();
831 		uvm_lock_pageq();
832 	}
833 #endif
834 
835 	/*
836 	 * now we want to work on meeting our targets.   first we work on our
837 	 * free target by converting inactive pages into free pages.  then
838 	 * we work on meeting our inactive target by converting active pages
839 	 * to inactive ones.
840 	 */
841 
842 	/*
843 	 * alternate starting queue between swap and object based on the
844 	 * low bit of uvmexp.pdrevs (which we bump by one each call).
845 	 */
846 	got_it = FALSE;
847 	pages_freed = uvmexp.pdfreed;	/* XXX - int */
848 	if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0)
849 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp);
850 	if (!got_it)
851 		got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj);
852 	if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0)
853 		(void) uvmpd_scan_inactive(&uvm.page_inactive_swp);
854 	pages_freed = uvmexp.pdfreed - pages_freed;
855 
856 	/*
857 	 * we have done the scan to get free pages.   now we work on meeting
858 	 * our inactive target.
859 	 */
860 	inactive_shortage = uvmexp.inactarg - uvmexp.inactive - BUFPAGES_INACT;
861 
862 	/*
863 	 * detect if we're not going to be able to page anything out
864 	 * until we free some swap resources from active pages.
865 	 */
866 	swap_shortage = 0;
867 	if (uvmexp.free < uvmexp.freetarg &&
868 	    uvmexp.swpginuse == uvmexp.swpages &&
869 	    uvmexp.swpgonly < uvmexp.swpages &&
870 	    pages_freed == 0) {
871 		swap_shortage = uvmexp.freetarg - uvmexp.free;
872 	}
873 
874 	for (p = TAILQ_FIRST(&uvm.page_active);
875 	     p != NULL && (inactive_shortage > 0 || swap_shortage > 0);
876 	     p = nextpg) {
877 		nextpg = TAILQ_NEXT(p, pageq);
878 
879 		/* skip this page if it's busy. */
880 		if (p->pg_flags & PG_BUSY)
881 			continue;
882 
883 		if (p->pg_flags & PQ_ANON)
884 			KASSERT(p->uanon != NULL);
885 		else
886 			KASSERT(p->uobject != NULL);
887 
888 		/*
889 		 * if there's a shortage of swap, free any swap allocated
890 		 * to this page so that other pages can be paged out.
891 		 */
892 		if (swap_shortage > 0) {
893 			if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) {
894 				uvm_swap_free(p->uanon->an_swslot, 1);
895 				p->uanon->an_swslot = 0;
896 				atomic_clearbits_int(&p->pg_flags, PG_CLEAN);
897 				swap_shortage--;
898 			}
899 			if (p->pg_flags & PQ_AOBJ) {
900 				int slot = uao_set_swslot(p->uobject,
901 					p->offset >> PAGE_SHIFT, 0);
902 				if (slot) {
903 					uvm_swap_free(slot, 1);
904 					atomic_clearbits_int(&p->pg_flags,
905 					    PG_CLEAN);
906 					swap_shortage--;
907 				}
908 			}
909 		}
910 
911 		/*
912 		 * deactivate this page if there's a shortage of
913 		 * inactive pages.
914 		 */
915 		if (inactive_shortage > 0) {
916 			pmap_page_protect(p, PROT_NONE);
917 			/* no need to check wire_count as pg is "active" */
918 			uvm_pagedeactivate(p);
919 			uvmexp.pddeact++;
920 			inactive_shortage--;
921 		}
922 	}
923 }
924 
925 #ifdef HIBERNATE
926 
927 /*
928  * uvmpd_drop: drop clean pages from list
929  */
930 void
931 uvmpd_drop(struct pglist *pglst)
932 {
933 	struct vm_page *p, *nextpg;
934 
935 	for (p = TAILQ_FIRST(pglst); p != NULL; p = nextpg) {
936 		nextpg = TAILQ_NEXT(p, pageq);
937 
938 		if (p->pg_flags & PQ_ANON || p->uobject == NULL)
939 			continue;
940 
941 		if (p->pg_flags & PG_BUSY)
942 			continue;
943 
944 		if (p->pg_flags & PG_CLEAN) {
945 			/*
946 			 * we now have the page queues locked.
947 			 * the page is not busy.   if the page is clean we
948 			 * can free it now and continue.
949 			 */
950 			if (p->pg_flags & PG_CLEAN) {
951 				if (p->pg_flags & PQ_SWAPBACKED) {
952 					/* this page now lives only in swap */
953 					uvmexp.swpgonly++;
954 				}
955 
956 				/* zap all mappings with pmap_page_protect... */
957 				pmap_page_protect(p, PROT_NONE);
958 				uvm_pagefree(p);
959 			}
960 		}
961 	}
962 }
963 
964 void
965 uvmpd_hibernate(void)
966 {
967 	uvm_lock_pageq();
968 
969 	uvmpd_drop(&uvm.page_inactive_swp);
970 	uvmpd_drop(&uvm.page_inactive_obj);
971 	uvmpd_drop(&uvm.page_active);
972 
973 	uvm_unlock_pageq();
974 }
975 
976 #endif
977